SfM photogrammetry uses computer vision algorithms for representing objects or scenes as 3D data and has become the main method used to derive topographic data from Unmanned Aerial Vehicle (UAV, or “drone”) photography. However, SfM has much broader application.
The Trinity River, California, was mined for gold for over a century, followed by construction of major dams that divert water and prevent flooding. This blocked the hydrologic ecosystem service of channel renewal and prevented channel recovery from mining, resulting in degraded in-stream and riparian habitats.
The Trinity River Restoration Program (TRRP) works toward returning ecosystem function through controlled flow releases and river channel rehabilitation. Rehabilitation is intended to be a beginning for the restoration of a site. Subsequent flows and vegetation growth are expected to complete the restoration, resulting in dynamic ecosystems. Expectations include recovery of vegetation on floodplains, migration and renewal of gravel bars, and more. Constructed log jams both provide shelter for young salmonids and interact with hydrologic forces during restoration flow releases.
TRRP has used SfM: to document channel rehabilitation features and detect changes with flows, to quantify tributary sediment contributions, and even to reconstruct the pre-dam topography and river channel from historical aerial photography.
TRRP has found that photographic needs for SfM projects range broadly. Camera requirements are minimal and allow opportunistic data when alternative methods such as laser scanning may not be practical; however, experience in photographic strategy for SfM is important. Modeling artificial log jams may require <50 images taken while moving around the jam, while side-channels and tributary deltas may require 500-2000 images and a complex strategy spanning the site with only small, incremental changes to the point of view between images.
Combined with survey-grade spatial control, models of complex terrain can span hundreds of meters with an internal model RMSE < 3 cm. However, internal model accuracy does not necessarily correspond to overall accuracy raw point clouds. SfM has some difficulty matching features across images, such as with repetitive patterns formed by dense branches, resulting in outlying groups of points that may require manual edits to the clouds.
Through SfM, tributary deltas were shown to have greater deposition during winter storms than previously thought, followed by transport during restoration flows. Changes to a constructed wood jam during a restoration flow were demonstrated to be less than suspected. SfM is proving to be a flexible method for quantifiably revealing ecosystem dynamics.